In-Context Forcing: Uncovering Context Effects in Autoregressive Video Diffusion

Autoregressive video diffusion models often struggle with a trade-off between temporal consistency and visual quality. Existing models typically rely on fully denoised previous frames as context, which leads to 'shortcut' learning—where the model simply copies local details from previous frames rather than generating coherent motion. This paper investigates how to provide better contextual guidance to improve temporal dynamics and inference speed.

The authors introduce In-Context Forcing, a paradigm that treats diffusion as a masking process. Instead of using clean frames as context, the model uses frames at varying noise levels. By applying higher noise to adjacent frames and lower noise to distant ones, the model receives adaptive guidance: coarse structural information early in the denoising process and fine-grained details later. To manage this during training, the authors propose a Step-wise Rolling KV Cache, which maintains these progressive contexts while ensuring that the model learns to handle the same noise-level distributions it will encounter during inference.

By decoupling the strict dependence on fully denoised frames, the proposed framework allows for cross-frame parallel denoising. This significantly accelerates inference speed compared to traditional sequential autoregressive models. Experiments on the VBench benchmark demonstrate that In-Context Forcing achieves superior visual fidelity and temporal consistency, effectively mitigating the error accumulation and detail leakage common in previous few-step autoregressive approaches.

This work addresses a critical bottleneck in real-time video generation. By enabling parallel denoising without sacrificing performance, the method makes high-quality, long-form video generation more feasible for interactive applications like game simulation and live content creation, where low latency is essential.